Towards Multi-Resource Physical Machine Provisioning for IaaS Clouds

Virtualization has been an enabling technology for IaaS (Infrastructure as a Service) Clouds. Physical machine (PM) provisioning is a key problem for IaaS cloud providers on their resource utilization and quality of service to users. Proper provisioning is able to ensure the service quality while conserving unnecessary power consumption from over-provisioned PMs. However, the effectiveness of PM provisioning in current IaaS providers such as Amazon and Rackspace is severely limited by that they offer virtual machines with proportional resource provisioning on different resource types (including CPU, memory and disk etc). Such a rigid offering cannot satisfy diversified user applications in the cloud, and can cause significant overprovision on PMs in order to satisfy users' requirement on all resource types. This paper argues a more flexible approach that IaaS providers should offer virtual machines with flexible combinations on multiple resource types. We further formulate the problem of multiple resource virtual machine allocations for IaaS clouds, and develop analytical models to predict the suitable number of PMs while satisfying a predefined quality-of-service requirement. Experiments show that the proposed approach can significantly increase the resource utilization, with a reduction on the number of active PMs by 27% on average.